Sparse Multi-way Digital Signal Processing Approach for Detection of Deep Medial Temporal Discharges from Scalp EEG

Lead Research Organisation: University of Surrey
Department Name: Computing Science


Detection of deep brain medial temporal discharges is extremely crucial for early detection of epilepsy and plan for a surgical operation to remove very small regions of the brain to avoid recurrent or progress of such very common neurological condition. Currently, this affects more than 1% of the UK population. In addition to clinical history, the scalp EEG is the most popular and accepted test to support the diagnosis of epilepsy. Unfortunately, such discharges (spikes) cannot be seen from the scalp EEG due to their relatively low sensitivity. 45% of awake EEGs and 20% of sleep EEGs recorded from people with epilepsy do not show clear epileptiform abnormalities. This leaves a significant proportion with uncertain diagnosis and delayed treatment. However, although many epileptiform discharges cannot be detected on the scalp EEG, they can be recorded using intracranial EEG electrodes implanted in deep brain structures. A non-invasive method for increasing the probability of detection of epileptiform discharges will be therefore vary crucial and valuable to increase the diagnostic power of scalp EEG.
Detection of deep brain discharges using scalp EEG recordings using advanced digital signal processing (DSP) hasn't been much explored in the literature.
In this proposal new algorithms will be developed to initially use a set of previously recorded data (in their so called training phase) to best model the neural pathways from deep medial temporal source to scalp potential patterns. Solving this problem, we can then perform separation of the weak spikes from noise-like scalp signals, and localize the sources. In this direction, the major problems are nonlinearity of the medium and interference of the cortical potentials which are usually recognised as the scalp EEG of a normal brain.
A large set of simultaneous scalp and intracranial EEG data using Foramen Ovale (FO) electrodes was collected from more than twenty patients have been recorded and analysed. Using some simple methods the signal-to-noise ratio (SNR) was increased by averaging the data over a number of trials synchronized on discharges using intracranial recording. It has been reported by providing many evidences that:
a) Before averaging only 9% of the discharges were detectable when only scalp recordings were used.
b) The majority of the spikes (up to 72.3%) could be detected by using both intracranial and scalp EEGs particularly after averaging.
c) In 18.7% of discharges no scalp signal was observed even after averaging.
From this unique and clinically important setup and the outcome of their analyses it is evident that interictal medial temporal epileptiform discharges, originating from deep medial temporal structure (MTS), can hardly be detected by visual inspection of the scalp signals. On the other hand, intracranial recording is a very inconvenient process and costly to the patients requiring hazardous and timely surgical operations.
From signal processing point of view, the deep sources are sparse, the medium is nonlinear, and the interfering signals are correlated and nonstationary. On the other hand, the number of sources is potentially larger than the number of electrodes which makes the overall system underdetermined.
In this proposal, to identify the underdetermined system when the sources are sparse and nonstationary, we will develop a sparse tensor factorization algorithm. The statistical (such as sparsity), geometrical (such as approximate locations), and physiological a priories (nature/shape of sources and artefacts) will be incorporated into the formulation as constraint terms. Finally, such a multiple constraint problem will be solved by developing a global optimization method.

Planned Impact

Detection of intermittent medial temporal discharges in their early stages is crucial for diagnosis, early treatment, patient monitoring, highly reducing the administration of anticonvulsant drugs, and consequently, reducing the side effects, and accuracy and punctuality in surgical planning leading to much less damage of the brain of the patients, particularly those suffering from epilepsy. Approximately 8% of the UK population suffer from epilepsy and the standard test to diagnose epilepsy is the EEG recorded from over the scalp. About 50% of patients with epilepsy show a normal EEG when awake and 80% show a normal EEG during sleep. Consequently, a method that could identify such deep discharges will be very crucial and clinically very important to confirm the diagnosis of epilepsy in 20%-50% of the patients with epilepsy.

The development during the course of this project, particularly the design of constrained sparse tensors and its associated mathematical and optimization solutions, will have much to offer to the signal processing community. It will influence multi-modal data analysis and information fusion at the presence of sparse events. The involved optimization problems and the provided solutions will have many applications in real life such as detection of abnormalities in foetal electrocardiography recorded from the mother abdomen, detection of rare substances from the hyper-spectral images, detection of fleet signals for navigation and military applications, and recovery of tampers from watermarked video sequences. Tensor factorization has attracted much attention in all areas of signal processing recently. It however, has much to improve in the future particularly in dealing with nonstationary data and nonlinear systems. Data fusion and multi-modal data analysis has yet to be researched. This concept has a major impact in places where multidimensional nonstationary data of different modalities are to be processed.

The advances in global optimization and solution for multiple constraints system will be a useful foundation for future research within the DSP, control, nonlinear systems, adaptive processes, and many other communities and applicable in many new systems with multi-modality measurements.

As direct impacts, the outcome of this project will first benefit a wide community of patients with interictal discharges who may develop focal epilepsy. Detection and localization of these interictal spikes will allow a minimally invasive operation to stop further development of such neurological disorder. The results will also alleviate the need for invasive operations to implant the FO electrodes within deep brain structure. This will decrease the risk of operation, time, cost, hospital ward space and the relevant resources for such surgical intervention and therefore, enhances both inward and outward clinical monitoring and treatments. The method therefore, would significantly help in early diagnosis of temporal lobe epilepsy. The early diagnosis is necessary for early treatment, which consequently would help in reducing seizures and in improving quality of life of patients with epilepsy and other patients with potential future seizure attacks.

Epilepsy Action is Britain's largest member-led epilepsy charity and acts as the voice for the UK's estimated 600,000 people with epilepsy, as well as their friends, families, carers, health professionals and the many other people on whose lives the condition has an impact. As a stakeholder in this study they help promote the research and disseminate findings to a lay audience. For example, by posting information on Epilepsy Action's website and publishing articles and updates in their magazines, Epilepsy Today and Epilepsy Professionals. They will also provide opportunities to present at local and national events. Therefore, during the course of the project we closely work with Epilepsy Action to enhance and strengthen this public relation.


10 25 50
Description We verified that by applying our signal processing and machine learning techniques to the data provided by KCL we can increase the possibility of detecting the temporal medial spikes from the scalp EEG signals. The increase can be from originally reported 9% to above 50%. Now we believe this is up to above 70%. In addition, the results of our work paved the path for (1) seizure prediction, and (2) that joint EEG-fMRI brain imaging system can be effectively and more accurately used for seizure.
Exploitation Route As the main objective of the project, the spikes should be extracted from the scalp EEG. This will help all clinicians working on seizure and epileptic seizure to diagnose the disease from the scalp EEG.

We managed to achieve detection of the spikes from scalp EEG up to 70%. Unlike what was previously achieved by the clinicians, i.e. only 9%, this will allow clinicians to avoid a significant number of pre-operation invasive assessments.

In addition, we introduced two theoretical concepts of (i) joint partial dictionary learning - blind source separation and (ii) coupled dictionary learning, both contribute to advances in knowledge in signal processing as well as machine learning. In addition, we are making the outcome useful for joint EEG-fMRI technique for the detection of deep epileptic brain discharges.
Sectors Education,Electronics,Healthcare,Manufacturing, including Industrial Biotechology

Description Two schemes have been developed for detection of the temporal medial discharges from the scalp electrodes in parallel. The first approach called constrained blind source extraction and dictionary learning exploits the information from intracranial wave forms to power up the proposed source separation technique. The second approach relies on machine learning approach for detection of epileptic discharges. In 2015 we expanded the research by employing combined machine learning and signal processing to increase the detection of medial temporal discharges from the scalp EEG. This allows detection of up to 70%. We also developed and applied new algorithms to better highlight and exploit inter subject variabilities by means of deep neural networks. There are 13 paper have been published out of this work and a new EPSRC proposal has been submitted in 2018. This will impact application of joint EEG-fMRI for seizure.
First Year Of Impact 2012
Sector Digital/Communication/Information Technologies (including Software),Education,Electronics,Healthcare
Impact Types Societal,Economic

Description In this project the collaboration is between Department of Computing, University of Surrey, and Department of clinical Neuroscience, King's College London. 
Organisation King's College London
Country United Kingdom 
Sector Academic/University 
PI Contribution Developing the necessary algorithms to be applied to patient data by our KCL collaborator.
Collaborator Contribution KCL has provided a large set of data for both the designing/training and testing stages of our proposed method.
Impact Provision of a very useful dataset to be used by Surrey Manual scoring/classification of data by KCL to be used by Surrey development of signal processing and machine learning algorithms developed by University of Surrey to be used by KCL.
Start Year 2013
Title Chirplet-based time-frequency modelling of the seizure spikes 
Description It allows to use a small number of chirplet bases to describe various types of seizure spikes. This helps to better (with less noise) model a limited number of spike wave forms to be used as templates in our proposed algorithms. 
Type Of Technology Software 
Year Produced 2014 
Impact This is later used to develop a dictionary of low noise templates. 
Description Group Meetings 
Form Of Engagement Activity A formal working group, expert panel or dialogue
Part Of Official Scheme? No
Geographic Reach National
Primary Audience Participants in your research and patient groups
Results and Impact We (the group in University of Surrey and the one in KCL) had our regular meetings 5 times since the beginning of the project. In these meetings all the investigators, two postdocs and two of the PhD students were attending.

This has stimulated the collaborations and the research members now or fully aware of their day to day duties.
Year(s) Of Engagement Activity 2014